The formula
Tokens per month: speed in tokens per second, multiplied by 2,592,000 seconds in a month and by the share of time the model actually generates. Price per million: the monthly costs from the local AI calculator, divided by the number of millions. Take the speed from measurements: for one user it is 19–145 tokens per second on models of 7–70 billion parameters. For batch output on a server card the total is, by our estimate, tens of times higher, and this is the main lever.
Examples at prices as of 29 September 2026
A Mac Studio M4 Max at $400 a month, a 7-billion model (Llama 2 7B in 4 bits), 83 tokens per second, generating 30% of the time: 64.6 million tokens and $6 per million, about $5.9. A workstation with an RTX 5090 at $800 a month, 128 tokens per second by the RTX 4090 measurement on an 8-billion model (the sources have no RTX 5090 measurement), 50% of the time: 165 million tokens, $5 per million. A 70-billion model in 4 bits takes about 38.5 GB and does not fit on one RTX 5090 with 32 GB. A workstation with two RTX 5090 cards costs about $1,100 a month, and at 19 tokens per second, by the measurement of two RTX 4090 cards, that is 24.8 million tokens and $46 per million, more than the output tokens of Opus 5.5 at $20. A rented H100 at $3 per hour on a 70-billion model at 25 tokens per second and 80% utilization: $41 per million; on an 8-billion model at 144.5 tokens per second, $7. A server with four H100 cards at $6,100 a month and batch output of 2,000 tokens per second in total (our assumption) at 60% utilization: 3.1 billion tokens and $2 per million, about $2.
What follows from this
One user on a local 7–8 billion model pays $5 to $6 per million tokens. That is the level of mid-range cloud models for output tokens, Haiku 4.5 at $5 and GPT-5.4 mini at $5, and 9–12 times more than cheap ones, GPT-6 Luna at about $0.50. Batch output for dozens of users on a server card brings the price down to $2 in the example with four H100 cards. That is cheaper than mid-range cloud models for output tokens, with full control over the data. A 70-billion model for one user costs $40 to $43 per million, like the cloud flagships: companies choose it for the data, not for the price.
How to measure your own speed
Rent the card you need for an hour: an RTX 4090 from about $0.50, an H100 from $3 per hour. Install an inference engine and run 50 of your own requests at 1, 4 and 16 parallel streams: the total speed at 16 streams is the number for the calculator. On a Mac, measure in LM Studio: it shows the speed after each answer.